Itay Z (@itayzit) header
Itay Z (@itayzit) profile picture

PM at Nvidia. I take long-form AI content and break it down. no fluff

San Francisco, CA
Joined November 2025
Photo by Itay Z (@itayzit)Photo by Itay Z (@itayzit)Photo by Itay Z (@itayzit)Photo by Itay Z (@itayzit)Photo by Itay Z (@itayzit)Photo by Itay Z (@itayzit)Photo by Itay Z (@itayzit)Photo by Itay Z (@itayzit)Photo by Itay Z (@itayzit)
The Navier Stokes saga taught us once again: the frontier of AI is spiky. @3blue1brown says that It's actually fractal! zoom into math, and it's spiky too. What is it not good at then? @3blue1brown splits it into 3 moves:
2
6
1
98
28,472
OpenAI ran agents to fix their CI, and they did. 1. Their prompt engineering is sooo trivial it's insane (see pic) 2. "the agent was most productive in the 5-day stretch that i didn't message it once" 3. "when the agent wasn't sure about something, it launched another agent to review its code"
OpenAI engineer Aaron Friel: "I've been promoted to a director of agents. I give an agent a job, and then it manages a team of agents to get it done for me" his Codex agent ran for 1,000+ hours fixing CI across all of OpenAI. last year, 7 hours felt like a big deal. the prompt was 200 words. zero code. over 6 weeks: - 100+ PRs a week - 213+ of its own fixes merged - 5,856 subagents - ~$32 an hour, ~$16 on the newer model it even messaged OpenAI researchers on its own to crack a GPU bug. its most productive stretch was the 5 days he didn't message it once. watch the 7-min story, then read below how I hired my first AI employee for my repos ↓
1
3
98
וואו, אנשים רוצים יותר נחליאלי בחיים שלהם מסתבר #bringnahlieliback
Replying to @omriasher
הפוך. המשקל שלו בחיים הבוגרים שלי לא גדול מספיק 🐦
1
56
pov: your chrome extension is being crawled by a romanian install-farm what to do??
1
1
104
rather than what??? source: github
3
122
let that sink in - the honest answer is that 'no fluff' is the most load-bearing phrase these days source: @hackernews
1
3
88
wow i love this. for the longest i’m feeling that turn based chat interfaces are so overused. who ever implements @liu_xingyu ‘s ideas in their product is gonna win so many consumers
AI interfaces nowadays are still pretty much turn-based. In my PhD thesis, I thought about what interaction would look like if we could go beyond taking turns. One example is intra-turn interaction:
2
162
“chat, make me a portfolio website”
openai's head of design's website omg iansilber.com/ 😭
70
hedging is beautiful 😍
Mammoth Rock ‘n’ Bowl is a restaurant and bowling alley in Mammoth, California, one of the top ski towns in America. Like the hotels, rentals, and shops around it, the business runs on winter tourism. A bad snow year hits revenue hard. Rock ‘n’ Bowl is especially exposed since they also sponsor a youth snowboarding team. So rather than wait and see, co-owner Ashley Marks asked Kalshi for a market on Mammoth Mountain December snowfall. We built it, and she’s trading $5,000 for a $31,250 payout if there’s weak snow. "We created the snow market with Kalshi because like other mom and pop businesses, we live with seasonal snow risk and wanted to give the community a way to offset some of that uncertainty." - Ashley If the snow comes, business runs like any other winter, and she loses the $5,000 trade. If it doesn’t, the Kalshi position pays out, giving Ashley $26,250 to help offset lost tourist revenue. Either way, Ashley's business is protected.
3
130
getting full agi scaries on the flight to sf 👻
6
316
osher ad is the new costco
האיירפודס במחיר של שבירת שוק
5
160
wow, first @trq212, now @_catwu, people are leaving right before the ipo @bcherny i guess you’re next
Friday was my last day at Anthropic. Two years ago, when Claude Code was a small internal tool, I asked it to build an RL env. I steered it for tens of hours but never manually wrote a line of code. It worked, I was hooked, and I joined the team. When we first launched, most people approved every permission prompt by hand and mostly used Claude to answer questions and write tests.  Today, the default is trusting Claude's judgment to auto-approve. Claude does large migrations and builds new products.  Helping build Claude Code into an industry-defining product has been the journey of a lifetime.  Thank you to the incredible team behind Claude Code, Cowork, and Tag. These products are the sum of thousands of great ideas from you. I'll miss working with you all. Thank you to our users for embracing a new way of working. Your constant feedback shaped so much of what we shipped and your stories made the work so meaningful. Excited to see what you all make next!
4
318
a new playbook emerges for companies that lost in the LLM race: 1. Become a services company 2. Go sovereign Examples: @MistralAI , @cohere , @AI21Labs
4
198
Fascinating result here. tl;dr: harnesses don't mean much. "Swapping the harness flips as many tasks as rerunning the same harness"
3
9
259
give a man a fish and you feed him for a day teach a man to Claude Code and you financially ruin him with distracting side projects for the rest of his life
416
1,671
203
25,753
643,281
Why train when astra/fable can just raw dog the weights directly
13
15
2
824
74,973
this week, OpenAI moves existing $200 subscribers to a lower usage allowance. their reason: "our increasingly efficient models". Are they basically charging for parallelization? 1. Fast mode uses your allowance at 2.5x the rate 2. Ultrafast uses it at 8x Last month @polynoamial said: on some benchmarks, 4 agents finish twice as fast, and you're paying 2x more to get an answer twice as quickly right??
1
1
80
long term, i think they’ll try to flip it towards subscriptions. right?
Replying to @socialcapital
In 1H 2026, ~80% of Anthropic’s revenue was consumption-based. The other ~20% was subscription-based.
1
2
310
Anthropic’s IPO will be the biggest day for philanthropy ever. $200B (!!) In Jan, The 7 founders pledged that 80% of their stake will be given to charity. At a $2T valuation, that will be the biggest donation to charity of all times. good luck
2
202